Qualitative and quantitative assessment of apple quality using bulk optical properties in combination with machine learning and chemometrics techniques

This study aimed to understand the quantitative relationship between the bulk optical properties (BOP), soluble solids content (SSC), and fruit firmness (FF) of apples, along with the qualitative discrimination of apple cultivar and shelf-life. The absorption coefficient ( μ a ) and reduced scattering coefficient ( μ s ′ ) of 200 apples from four cultivars during 36-days shelf-life were determined using the single integrating sphere technique in 500-1000 nm. Partial least squares regression (PLSR) and random forest (RF) algorithms were used to establish quantitative prediction models for SSC and FF based on the BOP of apples. The results indicated that the PLSR models based on μ ɑ and μ s ′ were optimal for quantitative prediction of SSC ( R 2 p =0.749, RMSEP=0.507) and FF ( R 2 p =0.745, RMSEP=0.571), respectively. RF and linear discriminant analysis (LDA) were used to establish qualitative models for discriminating apple cultivar and shelf-life, demonstrating that the RF model based on μ ɑ and μ ɑ + μ s ′ had the highest accuracy for the determination of apple cultivar and shelf-life, respectively, with the prediction set reaching 93.2 % and 85.7 %. Overall, RF was better than LDA for qualitative discrimination; however, it was less effective than PLSR for quantitative modeling. • BOP of apples in four cultivars during 36-days shelf-life were determined • Quantitative and qualitative evaluation of SSC, FF, cultivar and shelf-life were conducted • Absorption properties performed better for SSC prediction and cultivar identification • Scattering properties achieved better performance for modeling FF and shelf-life • RF outperformed LDA in qualitative discrimination, but was inferior to PLSR in quantitative modeling